AI is the next era, and I went all in before the market did.
I'll keep this plain. No jargon, no posturing, and no claim that I'm sharper than anyone on your side. I made the same bet and started a company on a simple idea: most small companies pay full price for a generic SaaS stack to use a tenth of it: subscriptions that never quite fit and quietly bleed money every month. Using nothing but AI, I build them a custom tool that does the job their way and replaces those subscriptions, so instead of renting software that's wrong for them, they own one that's right, and the recurring spend stops. A regulated shop was just the first real test of that, and it held. I'd rather show you the work and let you push on it.
- Built an offline tool for a regulated aerospace shop that cut estimating work about 50%, plus an AI support layer it reaches from outside its network. No AI runs inside the shop. All built through Claude Code.
- An AI hiring judge that has to cite real evidence, plus a verifier that catches it bluffing.
- A spread of other custom tools, shipped solo and full stack. More in the range below.
How I think and what I built. Not a plan for your company I have no business writing yet.
I replaced myself with the tool that did my old job. What one person can ship with AI is why they kept me on.
Worth being straight: that's one deployment, not a fleet. A clean record at n=1 is a start, not a statistic, and I'd rather say that than dress it up.
I started as an estimator. Then I built the estimating tool the shop runs today, end to end through Claude Code on a heavy spine of docs and specs. Here's a short walkthrough, and the real thing to poke at.
When I say regulated, I mean the paperwork holds up.
The customer-facing compliance documents that ship with the tool. Tap to read in place.
Pen-test results go to vetted procurement as a summary, not posted in full on a public page. Same judgment the whole tool is built on.
Two stories. I keep them separate, because the difference is the point.
- No AI inside it. The speed win is software, not a model.
- Runs fully offline, no telemetry.
- That offline design is exactly why a regulated shop can run it.
- Email and portal agents. Never installed in the shop.
- Built off-network, against aliased data, public vendor docs.
- Nothing AI ever touches the controlled environment, by design.
So how does a shop that can't host AI get AI help? On the far side of a boundary they already control.
- A separate agent for each function, each kept in its own lane so they never step on each other.
- Routed by urgency, with a person in the loop to confirm or escalate before anything real happens.
- Email and portal only. Nothing AI installed inside. An on-site agent add-on is roadmap, not shipped.
Pricing, kept separate: regulated license is $10K flat, unlimited seats. Off-the-shelf is ~$4,999/seat + $99/mo. One active license; a second on the table with a PE owner, not closed.
Don't take a savings number from me. Compute your own.
I built MFC's site and this model. Its rule: never claim a flat savings stat. Start with what an estimator costs you.
At the shop itself, the savings is the estimating department's, and it doesn't net out the pay for the role I moved into. While I'm there, that takes some of it back. The day I'm gone, the tool stays and the savings lands at 100%.
A non-coder edits the pricing engine in plain English. Nothing writes until a human says so.
Type a change in plain English. The real settings-editor AI reads it and shows you the exact before-and-after, so you see precisely what moves. Nothing is written until you Apply.
The text box is live: it calls the real settings-editor AI, the same prompt and rules the product uses, on a sample config. The example buttons are instant and run no model. Either way it works the same: the AI suggests, you see exactly what would change, and nothing is saved unless you say so.
Nick built a Devil's Advocate GPT. I built a judge that has to prove every claim it makes.
Screen drops a candidate into one messy business problem and watches how they solve it with AI: a blank build assistant, simulated stakeholders who each hold part of the picture, a gated expert you only reach if you ask, and a deliberate trap. Take it yourself.
- Scores against an anchored rubric, per dimension.
- Must quote verbatim evidence for every score.
- A verifier checks every quote, and catches fabricated evidence.
- No accuracy percentage. I haven't run a formal study.
- Reliability is a method, not a number.
- A Stage-0 prototype, no paying customers.
Admin view Inside the demo, the star button opens the reviewer view: the real judge run on three seeded runs, end to end.
- Per-dimension scores and the final band.
- The trap: caught, missed, or pre-empted.
- Citations: checked, fabricated, flagged.
Synthetic candidates, real scoring. The reviewer view runs the same anchored rubric and citation verifier as the live screen, and shows only genuine judge output, never hand-authored numbers.
Your Law 5 as running code: confident incorrectness comes standard, so fact-check often. The verifier literally fact-checks the grader.
The anchors aren't flukes. The spread:
- One person, full stack. Offline tool, backend, database, marketing site, admin CRM, live payments, all built through Claude Code. I design it, wire it up, and run the testing myself.
- Knowing what to stop. Built SpireMeta to 250 users in six weeks, then killed it on a cost call. Knowing what not to keep alive is the rarer skill.
- I trained the models first. At Pareto, out of 2,000+ trainers, I took on senior-project work and mentored double-digit cohorts. Prompt engineering and adversarial testing on projects for companies like Anthropic and Stanford.
- Trusted in the room. As Business Systems Developer through an active merger, I'm in weekly C-suite meetings and coordinating the systems integrations that connect both companies under one PE owner.
A conviction bet, in order.
Drive, curiosity, and perseverance are what the work takes. I was training frontier models after hours, before it was a category and before anyone asked me to.
Speed is strategy. I paused the degree a semester short to take the seat while the window was open. First-mover over the credential.
Adoption is won by whoever can make AI make sense to the people doing the work. A regulated shop floor is the hardest room for that, and it's the one I walked into.
AI is an amplifier, not a replacer. I went from grading the models to deploying one for a paying customer. Same tool, pointed at real work.
AI takes the boring parts, not the job. It took mine, and the shop kept me on to build what's next. Crawl, walk, run: that's already run.
The unlock is meeting people where they are, not dragging them into a new world. Getting AI adopted where it's resisted isn't a pivot for me. It's been the assignment all along.
MOD Pizza training manager (reviews up 20%+ in six months). Founder of Empty Ego ($500 to six figures over 22 months). Sales and member services at Sunrun and UnitedHealth.
Before the calculator: Excel, HTML, docs, ERP add-ons. Then the realization that an offline Electron app was the right shape, which became MFC.
An event-inventory tool and a site-builder. Formed Psyrcuit LLC around Nov 2025. Productized MFC, and built the stats site I later killed.
The shop sold to a PE owner. As Business Systems Developer I'm now in weekly C-suite meetings through the merger, coordinating the systems integrations to bring both companies under one owner. Through that same PE, a second finishing company is in discussion for the same $10K structure. On the table, not closed.
By any metric, I'm already running.
Agentic AI in production, a human in the loop, in real environments. And you just poked at it yourself. I'm not telling you I could get there. I'm already there.
All of it is pokeable on purpose. If a claim didn't survive you pushing on it, tell me. That's the standard I hold my own work to.
I'm Chase Lance. I ship production AI where it's hard to ship. If that's who you want in the room, that's where I'd like to be.